Broadcom sees AI chip boom as it challenges Nvidia

Nvidia still owns the AI chip conversation, but Broadcom is quietly building a business that makes the question more complicated than it used to be.

Broadcom reported $16.7 billion in AI semiconductor revenue for its fiscal third quarter of 2026, a 221% jump from the same period a year earlier.

The custom chip strategy paying off

The key to understanding Broadcom’s growth is the distinction between what it sells and what Nvidia sells. Nvidia makes powerful general-purpose graphics processing units that can run almost any AI workload out of the box. Broadcom makes custom application-specific integrated circuits, or ASICs, designed from scratch around a specific customer’s exact needs.

Those custom accelerators accounted for 73% of Broadcom’s Q3 AI revenue, and the client list reads like a who’s who of frontier AI development. Google has partnered with Broadcom across multiple generations of its Tensor Processing Units. Meta’s MTIA chip family runs through the same pipeline. OpenAI is developing its so-called Jalapeño chips with Broadcom’s help, and Anthropic has joined the roster as well.

Broadcom claims a roughly 70-80% market share in custom AI accelerator design.

Forecasts that raise eyebrows

Management revised its fiscal 2026 AI revenue guidance upward to approximately $58 billion, representing roughly 186% growth year-over-year. From there, the projections accelerate sharply: fiscal 2027 guidance targets around $115 billion, up from a prior estimate of more than $100 billion, and fiscal 2028 carries a forecast of approximately $230 billion.

Broadcom ties that projection to what it describes as multi-gigawatt deployments, meaning data centers so large they require their own power infrastructure.

The total AI backlog, based on current demand and active projects, exceeds $73 billion.

Where Broadcom fits in the broader AI hardware picture

Nvidia dominates the training of large AI models, where raw compute flexibility matters most. Broadcom thrives in inference at scale, where a company like Google knows precisely what computation it needs to perform billions of times per day and can justify the engineering investment to build a chip optimized for exactly that task.

The companies also overlap in AI networking silicon, where Broadcom has a long-established position in switching and connectivity chips that move data between processors inside large clusters.

Designing a custom chip is a multi-year collaboration. Google has been working with Broadcom on TPUs across multiple chip generations, which means the switching costs are genuinely high and the revenue is sticky in a way that commodity hardware revenue rarely is.

Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our Editorial Policy.